/MLG/- Machine Learning General

what are you working on/studying?

>learning - youtube.com/watch?v=UzxYlbK2c7E&list=PLA89DCFA6ADACE599
youtube.com/watch?v=mbyG85GZ0PI&list=PLD63A284B7615313A
youtube.com/watch?v=_PwhiWxHK8o&list=PLnvKubj2-I2LhIibS8TOGC42xsD3-liux

>statistics - youtube.com/watch?v=KbB0FjPg0mw&list=PL2SOU6wwxB0uwwH80KTQ6ht66KWxbzTIo

>competitions - kaggle.com/

Other urls found in this thread:

papers.nips.cc/paper/6672-unsupervised-image-to-image-translation-networks.pdf
make.girls.moe/
web.stanford.edu/~jurafsky/slp3/9.pdf
cs231n.github.io/
my.mixtape.moe/kiyaun.webm
twitter.com/NSFWRedditImage

Nothing to be honest.
I am using some outdated version of openCv for work. Which of the frameworks in Op's image would play the nicest with my existing C/C++ code?

idk any framework that uses c/c++ sorry. a lot of them use python my guess is because of numpy and scipy which makes it comparable to matlab when manipulating tensors.

Tensorflow, no question

Is there a good ML framework for C#?

Your'e doing God's work OP.

>question for this thread
what /lit/ would /mlg/ recommend for getting started in ml?
Furthermore, what would be the best lang for ml?

can you give your background? math wise and programming wise so i can try to better answer your question? unless you were talking in general.

I am talking in general.
The most complex AI i've programmed so far was a brute-force maze navigator.

Elements of Statistical Learning if you have decent math skills. If not Introduction to Statistical Learning. This will give you a decent base to start and the tools to work with beyond being a simple code monkey.

>elements of statistical learning
who is the author?
is the full name "The elements of statistical learning Data mining, inference, and prediction"?

can someone more educated than me make a network that receives clothed and outputs nude?

you could use 3d models as initial training data and work your way up I guess

i'm currently reading "all of statistics" by larry wasserman. there are lectures online but judging from the first two they're pretty shit.
i started with an edx course mit's 15.071x and found it to be quite useful in introducing concepts
introduction to statistical learning is good.
hastie, tibshirani, friedman

We have tried so many times to have a machine learning general.

We have fail everytime.

I think we can work something out

papers.nips.cc/paper/6672-unsupervised-image-to-image-translation-networks.pdf

Yea thats it and this guy has the authors right If you can work through the book it will give pretty much all the background you need. I have a copy for reference and have gone through it a few times.

using nets for data augmentation, interesting.

not this time.
Quick, any anime waifus for the general?

We have unlimited waifus user:

make.girls.moe/

That's even better than what I thought!, thank you.

What languages are you all using?
I find R the best for EDA, but then generally end up doing everything ML related in python.
Right now I'm working on some time series stuff. Just a personal project trying to find indicators which predict regime shifts. Getting nice results with HMMs.

Can you explain HMM's in layman terms?

Tay would be a great start for a chan-bot waifu.

Also,
>machine generated waifu to represent Sup Forumsmlg/
sounds fucking perfect.

Post 'em /mlg/, lets see if we can find ourselves a thread mascot/waifu.

Trying to train a standard ResNet50 with anime pictures from danbooru. Unfortunately I kind of lost access to my training machine and now I need to find another computer with sufficient performance.

training takes so long i want to kill myself trying to train small nets on my x220

In layman's terms is pretty difficult, but I'll give a shot. It's actually pretty difficult to describe without math as well.

The basic idea is that we have a mixture model which has observable parameters and underlying parameters which aren't visible. These underlying parameters make up the mixture component and are actually not independent of each other and are part of a Markov process. From the observed parameters we can infer the state probabilities of the Markov process.

why not use this opportunity to figure out how to make the training more efficient and thus less system-costly?

yeah thanks, this does a good job explaining web.stanford.edu/~jurafsky/slp3/9.pdf

what exactly is a regime shift?

A regime shift is when the time series moves from acting with one formula to another. It could be increased instability, change in direction, change in intensity, etc.

If you have millions of pictures then just loading them into GPU memory takes its time. Even if you use some optimization tricks to reduce the number of training epochs its still necessary that the network looks at each image a few times…

The compressed (!) images don't even fit into normal RAM so I have no idea how to speed this up more on the GPU.

what if you fragment the image and feed the fragmented pieces in order to the machine, thus allowing it to process it in smaller chunks.

Even better, if it works to your advantage you could train the machine on the fragmented images themself.

Another trick would be to reprogram in assembly and directly control the processing o the image by the machine.

Just be creative user, there's more then one way to fry an egg.

>Even better, if it works to your advantage you could train the machine on the fragmented images themself.
Pretty sure if you did this, and generated fragments. You would need to encode that into your loss function of how well they matched and that would be way more work and computation.

>Another trick would be to reprogram in assembly and directly control the processing o the image by the machine.
I would hardly call myself a programmer, lets say programmer in training with a math background.

There is a Stanford online course based on these books at lagunita.Stanford.edu called statistical Learning. It's a good introduction and has examples in R. Alsoyou can find some resources from their regular courses like cs231n.github.io/

You know what? You're right! While thinking what an ignorant prick you are I had some interesting ideas.

I don't think smaller chunks and assembly will solve the problem because the bottleneck currently seems to be slow memory.

I have around 2 million images in webp format at the moment and the pipeline is:

HDD -> RAM -> GPU
Over and over again because my RAM is not sufficient. The HDD access really kills me here.

The boring solution would be to just use more aggressive webp compression but then the images will look like shit. But maybe it would be also possible to prepare the mini-batches beforehand and compress these. Instead of 2 million files there will only be 2 Mio/256 files (although larger) or something like this (depending on the batch size). I just have to look into lossy compression more. Maybe I can just use the montage tool from GraphicsMagick here.

OK. Can you elaborate why? I'm interested and you seem very confident...

I am doing it more as a personal project and a refresher since I haven't done much with Time Series since finishing my MS. Also trying out some different ideas from some other papers I've read, but the HMMs are giving the best indicators.

I have some ancient openCv could that has the same problems, I guess. It is written in Delphi and I ported it to C++. I thought I could speed it up when I parallelize it, but on a closer look there are so much unnecessary IO operations due to compatibility with Legacy code, that I identified that as the problem. I have already a speed up of ca. 20 min because of the rewrite (it was really bad code)

desu i was talking out of my ass for some of those suggestions. Apologies for that.

whatever idea you just came up with, i like it!

>I would hardly call myself a programmer, lets say programmer in training with a math background.
Mathematics is the basis for all CS. If you know math you have the key to it. The only part left is being creative in how you use and abuse that math to find solutions.

it might be the basis, but a lot of programming is knowing what you're doing and implementing. anyone that can think can program, but being good doesn't really correlate to having a deep mathematical background. statistics and machine learning share more overlap, but even then you don't go very deep.

>updating nvidia drivers before reinstalling cuda
>shut down the computer before driver gets installed like an idiot
>get "Failed to Create Kernel Channel 22" error on startup
>can't access any of my hentai ai decensorship work
Thankfully fixed my mistake by repairing packages in recovery mode.
No updates. Theano/Lasagne is a bitch to work with since the community is dead now.
For those who haven't seen this already, here's a (NSFW) visualization of the space of images my hentai neural network has learned to generate: my.mixtape.moe/kiyaun.webm